CUDA graphs

CUDA graphs#

For fixed task sets and timesteps, DLSSolver can capture a one-step solve_and_integrate graph to reduce launch overhead.

Requirements#

  • CUDA device

  • Fixed task list (same objects each frame)

  • Fixed dt and damping between captures

  • No host→device copies inside the captured region

Usage#

solver = mw.DLSSolver(cfg)
cfg.set_integration_dt(0.01)  # written before capture

# Warm-up (allocates task buffers, compiles kernels)
solver.solve_and_integrate(tasks, dt=0.01, use_graph=True)

while running:
    solver.solve_and_integrate(tasks, dt=0.01, use_graph=True)

If tasks or dt change, call invalidate_graph() (or recreate the solver).

Implementation notes#

  • Integration uses out-of-place qpos writes; in-place aliasing breaks graphs.

  • dt is stored in a device buffer via set_integration_dt before capture.

  • LM / L-BFGS / constrained backends do not expose graph capture yet.